Skip to content
Interview Prep Buddy LogoInterview Prep Buddy

Phase 4: Multiple Choice Questions

Phase 4: Multiple Choice Questions

Section titled “Phase 4: Multiple Choice Questions”

Test your knowledge with 25+ MCQs across easy, medium, and hard difficulty levels.

Easy (10 Questions)

Section titled “Easy (10 Questions)”

1. What does LLM stand for?

  • A. Large Logic Machine
  • B. Large Language Model
  • C. Linear Learning Module
  • D. Latent Language Mechanism
Show Answer **B. Large Language Model**

2. Which company created GPT-4?

  • A. Google
  • B. Anthropic
  • C. OpenAI
  • D. Meta
Show Answer **C. OpenAI**

3. What is the primary input format for an LLM?

  • A. Images
  • B. Audio
  • C. Text
  • D. Video
Show Answer **C. Text** (though many modern LLMs are multimodal)

4. What happens when an LLM receives a prompt longer than its context window?

  • A. The model crashes
  • B. The prompt is truncated
  • C. The context window expands automatically
  • D. The output is doubled
Show Answer **B. The prompt is truncated** (oldest tokens are dropped)

5. Which decoding strategy is completely deterministic?

  • A. Top-k sampling
  • B. Top-p sampling
  • C. Greedy decoding
  • D. Beam search
Show Answer **C. Greedy decoding**

6. What is a token in the context of LLMs?

  • A. A security credential
  • B. A unit of text (word or subword)
  • C. A type of neural network layer
  • D. A training hyperparameter
Show Answer **B. A unit of text (word or subword)**

7. Which model family does Claude belong to?

  • A. OpenAI
  • B. Anthropic
  • C. Google DeepMind
  • D. Meta
Show Answer **B. Anthropic**

8. What is the main purpose of tokenization?

  • A. To encrypt the input
  • B. To convert text to numerical IDs
  • C. To compress the data
  • D. To translate languages
Show Answer **B. To convert text to numerical IDs**

9. Which neural network architecture do modern LLMs use?

  • A. RNN
  • B. CNN
  • C. Transformer
  • D. GAN
Show Answer **C. Transformer**

10. What does SFT stand for?

  • A. Special Fine-Tuning
  • B. Supervised Fine-Tuning
  • C. Sequential Feedback Training
  • D. Simplified Forward Transform
Show Answer **B. Supervised Fine-Tuning**

Medium (10 Questions)

Section titled “Medium (10 Questions)”

11. What problem did the Transformer solve that RNNs couldn’t?

  • A. Handling variable-length input
  • B. Parallelization during training
  • C. Processing text data
  • D. Using neural networks
Show Answer **B. Parallelization during training** — RNNs process sequentially, Transformers process all tokens in parallel

12. In self-attention, what does the Query vector represent?

  • A. The information the token carries
  • B. What the token is looking for
  • C. The token’s position in the sequence
  • D. The final output of the attention layer
Show Answer **B. What the token is looking for**

13. Why do Transformers need positional encoding?

  • A. To make the model faster
  • B. Self-attention is permutation-invariant (order-agnostic)
  • C. To reduce memory usage
  • D. To enable parallel computation
Show Answer **B. Self-attention treats inputs as sets, not sequences — it doesn't know order without positional encoding**

14. What happens when temperature is set to 0?

  • A. Output becomes random
  • B. Output becomes deterministic
  • C. Output doubles in length
  • D. Model crashes
Show Answer **B. Output becomes deterministic** — the softmax approximates argmax, always picking the most likely token

15. How does DPO differ from RLHF?

  • A. DPO is slower but more accurate
  • B. DPO doesn’t need a separate reward model
  • C. DPO requires more human data
  • D. DPO uses supervised learning only
Show Answer **B. DPO directly optimizes preferences without training a separate reward model**

16. What is the “alignment tax”?

  • A. The cost of hiring human annotators
  • B. A slight performance decrease after alignment
  • C. GPU compute cost for RLHF
  • D. Legal compliance costs
Show Answer **B. Alignment can slightly reduce output diversity and creativity**

17. How does streaming improve user experience?

  • A. Reduces total generation time
  • B. Reduces perceived latency
  • C. Improves output quality
  • D. Reduces cost
Show Answer **B. Streaming sends tokens as they're generated, so users see output sooner**

18. Which of the following is NOT a known LLM family?

  • A. Gemini
  • B. Llama
  • C. Cortex
  • D. Mistral
Show Answer **C. Cortex** — this is not a major LLM family

19. What is the typical dimension of each attention head (d_k) in modern transformers?

  • A. 32
  • B. 64-128
  • C. 512
  • D. 1024
Show Answer **B. 64-128** — d_k = d_model / num_heads, typically 64-128

20. Why does RLHF use a separate reward model?

  • A. To speed up training
  • B. To provide a signal for optimizing the LLM
  • C. To reduce memory usage
  • D. To generate training data
Show Answer **B. The reward model approximates human preferences and provides a gradient signal for PPO optimization**

Hard (5 Questions)

Section titled “Hard (5 Questions)”

21. A transformer has d_model=4096, num_heads=32. What is the dimension of each attention head?

  • A. 64
  • B. 128
  • C. 256
  • D. 512
Show Answer **B. 128** — d_k = 4096 / 32 = 128

22. What is the relationship between the feed-forward network’s inner dimension and d_model in GPT?

  • A. Same as d_model
  • B. 2x d_model
  • C. 4x d_model
  • D. 8x d_model
Show Answer **C. 4x d_model** — The FFN typically expands to 4x the hidden dimension then projects back

23. Which technique constrains LLM generation to produce valid JSON?

  • A. Temperature scaling
  • B. Constrained decoding / grammar-based generation
  • C. Beam search
  • D. Top-k sampling
Show Answer **B. Constrained decoding uses a grammar to ensure output follows a valid JSON schema**

24. What is the Chinchilla optimal token-to-parameter ratio?

  • A. 10 tokens per parameter
  • B. 20 tokens per parameter
  • C. 50 tokens per parameter
  • D. 100 tokens per parameter
Show Answer **B. 20 tokens per parameter** — DeepMind's Chinchilla paper showed this is compute-optimal

25. In the original Transformer paper, how many encoder and decoder layers were used?

  • A. 6 encoder, 6 decoder
  • B. 12 encoder, 6 decoder
  • C. 6 encoder, 12 decoder
  • D. 8 encoder, 8 decoder
Show Answer **A. 6 encoder and 6 decoder layers** in the base Transformer model

Score Interpretation

Section titled “Score Interpretation”
ScoreLevelNext Steps
0-10🟢 BeginnerReview Module 1-2 basics
11-18🟡 IntermediateFocus on Modules 3-4
19-22🟠 AdvancedReady for Phase 5
23-25🔴 ExpertInterview ready!